Oncofusion

start / graph / ana-q163-forty-two-percent-of-the-occupied-background-is-wild-type-mll1-and-the-class-is-not

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

42% of every gene-body occupancy call is at a locus where wild-type MLL1 is also bound - but the AFF1-exclusive class is not, HOXA9 and MEIS1 are not, and SPAG7 is

also called q163 N and C overlap · the occupied background is 42% wild-type · SPAG7 is a wild-type MLL1 site · IRX1 is C-negative in 8 of 8

The antibody this project relies on recognises both the fusion protein and the normal version of the same protein. Comparing it with a second antibody that sees only the normal version shows that about four in ten of the sites counted as 'occupied' have the normal protein there too. The handful of genes the project cares about are not among them - with one exception, which should be dropped.

The same thing, for a scientist

q162 flagged that the median N/C ratio across all occupied loci is **1.77**, and asked whether what this project calls occupancy is largely **wild-type MLL1**. A ratio cannot answer that. **A set overlap can, and nobody had computed it**: apply the identical rule to the KMT2A-C arm - whose epitope is **not retained in any KMT2A fusion** - and ask how much of the N-occupied set it recovers. q154 reported the two pass rates separately (N 0.608%, C 0.469%) but **two sets of similar size can be identical or disjoint** [E1]. 🔴 **Pooled across 27 samples, 1,258 of 2,980 N-occupied calls (42.2%) are also C-occupied** - median **43.3%** per sample, range **8.8% to 74.2%**, and **92-fold over chance**. **So roughly four in ten gene-body occupancy calls this project makes are at loci where wild-type MLL1 is demonstrably bound**, and the overlap is nowhere near random. That is a standing caveat on the instrument, and it is the answer to the question q162 raised [E1]. 🟢 **But the genes that matter are on the other side of that line.** Scoring the C arm at the six class members across the 8 KMT2A::AFF1 samples [E2]: | gene | N-occupied | **C-occupied** | reading | |---|---:|---:|---| | PAN3 | 7 of 8 | **0 of 8** | C-negative | | ELF1 | 5 of 8 | **0 of 8** | C-negative | | GNAQ | 5 of 8 | **0 of 8** | C-negative | | **IRX1** | 3 of 8 | **0 of 8** | **C-negative** | | H3C15 | 3 of 8 | **1 of 8** | mostly C-negative | | **SPAG7** | 3 of 8 | **3 of 8** | **C-occupied wherever it is N-occupied** | | *HOXA9* (control) | 3 of 8 | **0 of 8** | C-negative | | *MEIS1* (control) | 2 of 8 | **0 of 8** | C-negative | 🔵 **The controls validate the readout.** `HOXA9` and `MEIS1`, the canonical KMT2A-fusion targets, are **C-negative in 8 of 8** - which is the signature a fusion-bound locus should give, and it was not assumed, it was measured [E2]. 🔴 **SPAG7 should be removed from the class.** It is **C-occupied in 3 of 3** samples where it is N-occupied. Two independent measurements now agree: q162 put its N/C at **0.49x the occupied background** - below, not above - and here it is the only class member the wild-type-only antibody finds every time. **SPAG7 is a wild-type MLL1 site that the N arm cannot distinguish from a fusion site** [E2]. 🟢 **The leading alternative explanation was tested and FAILED, which strengthens the result.** Histone clusters are 31% of all calls and are short, repetitive and highly transcribed - exactly the loci both arms might co-call for non-biological reasons. If they supplied the overlap, the rest of the call set would be clean. **They do not: histone loci overlap at 46.8% against 40.1% for everything else - only 1.17x, supplying 34.7% of the overlap while being 31.3% of the calls, and paired across samples the sign test gives p = 1.00.** So **40% of NON-histone gene-body calls are also wild-type-bound**. The attribution problem is general, not confined to a suspicious locus class [E3]. 🟠 **And the class has now been eroded from every side.** ELF1, GNAQ and PAN3 are refuted as non-exclusive by the authors' own caller; H3C15 carries three independent strikes; SPAG7 is wild-type. **`IRX1` is the only member that survives all four tests** - exclusive in an independent caller, above background on N/C, C-negative in 8 of 8, and the only one with a clean margin. **A class of one is not a class**, and that is the honest headline: what this project has is a single locus, not a partner-specific programme.

Go deeper
What the 42% does and does not mean

🔵 It does not mean 42% of calls are artefacts. A C-arm call means wild-type MLL1 is bound there; the fusion may be bound there too, and at a locus both proteins occupy the N arm is reporting something real. The correct reading is that 42% of calls are UNATTRIBUTABLE between fusion and wild-type, which is a different and more precise complaint than 'background'.

🟠 It is not uniform, and the spread is unexplained. Per-sample overlap runs from 8.8% (TB995291) to 74.2% (A73051). AFF1 samples sit slightly lower than the rest (median 38.4% against 45.2%), which is the direction fusion-specific signal would push, but the difference is small against that spread and no test was run on it.

🔴 The obvious alternative was checked and does not dissolve it. If the C arm were simply insensitive, C-negativity would be uninformative - but the C arm's own pass rate is 0.469% against the N arm's 0.608%, three-quarters as many calls genome-wide, and the overlap it does find is 92-fold over chance. A weak antibody does not produce a 92-fold-enriched overlap and then miss HOXA9, MEIS1 and IRX1 in 8 of 8.

🟠 What is still not excluded: that C-negativity at a specific locus reflects epitope accessibility rather than the absence of wild-type MLL1. Nothing here rules that out for any individual gene; it is the consistency across HOXA9, MEIS1 and four class members that carries the argument, not any single call.

Facts and figures
scriptanalyses/q163-is-the-occupied-background-wild-type-mll1/run.py
statusrun
answersthe question q162 raised about the occupied background's N/C of 1.77
spag7 verdictremove from the class - C-occupied in 3 of 3 samples where N-occupied, and q162 independently put its N/C at 0.49x the occupied background
pooled
pct42.2pct: 42.2also_C_occupied1,258also_C_occupied: 1258n_occupied_calls2,980n_occupied_calls: 2980
class members C arm in 8 AFF1
IRX13.000IRX1: 3H3C153.000H3C15: 3SPAG73.000SPAG7: 3ELF15.000ELF1: 5GNAQ5.000GNAQ: 5PAN37.000PAN3: 7
pooled — the numbers behind the chart
{
  "n_occupied_calls": 2980,
  "also_C_occupied": 1258,
  "pct": 42.2
}
per sample
{
  "median_pct": 43.3,
  "min_pct": 8.8,
  "max_pct": 74.2,
  "min_sample": "TB995291",
  "max_sample": "A73051",
  "median_fold_over_chance": 92,
  "aff1_median_pct": 38.4,
  "other_median_pct": 45.2
}
arm pass rates
{
  "N": 0.608,
  "C": 0.469,
  "unit": "percent of 19,402 gene bodies"
}
class members C arm in 8 AFF1 — the numbers behind the chart
{
  "PAN3": {
    "N": 7,
    "C": 0
  },
  "ELF1": {
    "N": 5,
    "C": 0
  },
  "GNAQ": {
    "N": 5,
    "C": 0
  },
  "IRX1": {
    "N": 3,
    "C": 0
  },
  "H3C15": {
    "N": 3,
    "C": 1
  },
  "SPAG7": {
    "N": 3,
    "C": 3
  }
}
positive controls
{
  "HOXA9": {
    "N": 3,
    "C": 0
  },
  "MEIS1": {
    "N": 2,
    "C": 0
  },
  "note": "canonical fusion targets are C-negative, validating the readout"
}
class status after five cycles
{
  "ELF1": "refuted - non-exclusive in the authors' caller",
  "GNAQ": "refuted - non-exclusive in the authors' caller",
  "PAN3": "refuted - non-exclusive in the authors' caller",
  "H3C15": "three strikes - histone cluster, worst margin 1.01x, N/C at background",
  "SPAG7": "refuted - wild-type MLL1 site",
  "IRX1": "survives all four tests; a class of one is not a class"
}
stratified by locus class
{
  "histone": {
    "both": 437,
    "N": 934,
    "pct": 46.8
  },
  "other": {
    "both": 821,
    "N": 2046,
    "pct": 40.1
  },
  "histone_over_other_fold": 1.17,
  "histone_share_of_overlap_pct": 34.7,
  "histone_share_of_calls_pct": 31.3,
  "paired_sign_test": {
    "histone_higher_in": 13,
    "n": 26,
    "p_two_sided": 1
  },
  "verdict": "the histone explanation is REFUTED; 40.1% of non-histone calls are also wild-type-bound"
}
measurements
[
  {
    "quantity": "gene-body occupancy calls that are also called on the wild-type-only C arm",
    "value": 42.2,
    "unit": "percent (1,258 of 2,980 pooled over 27 samples)",
    "subject": "the KMT2A-C epitope is not retained in any KMT2A fusion, so these calls are unattributable between fusion and wild-type MLL1",
    "universe": "median 43.3% per sample, range 8.8-74.2%, 92-fold over chance; C-arm pass rate 0.469% against the N arm's 0.608%"
  },
  {
    "quantity": "class members that are C-occupied wherever they are N-occupied",
    "value": 1,
    "unit": "of 6 (SPAG7, 3 of 3)",
    "subject": "IRX1, PAN3, ELF1 and GNAQ are C-occupied in 0 of 8; H3C15 in 1 of 8; HOXA9 and MEIS1 controls in 0 of 8",
    "universe": "each based on the 8 KMT2A::AFF1 samples; SPAG7's and IRX1's calls both rest on n = 3 N-occupied samples"
  }
]
Still open
  1. Test the next alternative explanation, now that the histone one is dead. Locus length, expression level and mappability were not controlled within the non-histone set, and a shared artefact in two antibodies run on the same chromatin with the same normalisation would produce exactly this pattern.
  2. Formally remove SPAG7 from the six-gene class and restate the class as what it now is. Two independent measurements agree it is a wild-type MLL1 site.
  3. Ask whether IRX1's survival is an artefact of serial elimination. Five members have been removed one at a time by five different criteria; applying each criterion to IRX1 with the same severity is the honest test, and the n = 3 denominator is its weakest point.
  4. Check whether the 8.8% to 74.2% per-sample spread tracks anything - depth, purity, proliferation. A three-fold range in how much of the call set is wild-type-attributable is itself a per-sample quality metric nobody has.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Applying the identical gene-body occupancy rule to the KMT2A-C arm, whose epitope is not retained in any KMT2A fusion, and intersecting with the KMT2A-N occupied set: pooled across 27 GSE252378 samples, 1,258 of 2,980 N-occupied calls (42.2%) are also C-occupied, with a per-sample median of 43.3% (range 8.8% to 74.2%) and a median 92-fold enrichment over the chance overlap expected from a random C-set of the observed size; KMT2A::AFF1 samples give a median of 38.4% against 45.2% for the other partners. The C arm's own genome-wide pass rate is 0.469% against the N arm's 0.608%.

    PMID 39472576DOI
    What this evidence cannot show

    The chance baseline is the marginal expectation from a random C-set of the observed size and does not hold the coverage or length structure of the gene-body table fixed, so the 92-fold figure overstates enrichment to the extent that both arms preferentially call the same short, highly-covered loci - and q161 showed 31% of calls are histone clusters, which are exactly such loci. The 42.2% pooled figure weights samples by call count. Only the flat-floor variant of the rule was run. Both arms share the same normalisation, so a normalisation artefact would appear in both and inflate the overlap.

    Evidence that argues against this (2)
    • The overlap could be driven almost entirely by the histone clusters rather than distributed across the call set: if the ~31% of calls that are histone-cluster genes are co-called by both arms, they would supply most of the 42% on their own, and the remaining calls could be largely C-negative. This analysis did not stratify the overlap by locus class, and it should have.
    • A C-arm call does not establish that the fusion is ABSENT - only that wild-type MLL1 is present. At a locus both occupy, calling the N-arm signal 'unattributable' is correct, but calling it contaminated would not be. The 42% is a ceiling on how much of the call set could be wild-type-only, not an estimate of it.
  2. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Stratifying the N-and-C overlap by locus class refutes the hypothesis that it is carried by histone clusters: pooled across 27 samples, 437 of 934 occupied histone-cluster calls (46.8%) are also C-occupied against 821 of 2,046 occupied non-histone calls (40.1%), a ratio of only 1.17; histone loci supply 34.7% of the total overlap while constituting 31.3% of the calls, and paired across samples the histone overlap is higher in 13 of 26, two-sided sign test p = 1.00.

    PMID 39472576DOI
    What this evidence cannot show

    This tests one alternative explanation and rejects it; it does not test others. Locus length, expression level and mappability were not controlled for within the non-histone set, and a different high-coverage locus class could still supply a disproportionate share. Histone-cluster membership is a symbol-prefix regex.

    Evidence that argues against this (1)
    • Rejecting the histone explanation makes the 40% harder to dismiss but does not establish that those loci are genuinely wild-type-bound rather than co-artefactual in both arms. Two antibodies applied to the same chromatin with the same normalisation share every upstream failure mode, and a shared artefact would produce exactly this pattern.
  3. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Scored on the KMT2A-C arm across the 8 KMT2A::AFF1 samples, the six members of the gene-body AFF1-exclusive class divide sharply: PAN3 (N-occupied in 7 of 8), ELF1 (5 of 8), GNAQ (5 of 8) and IRX1 (3 of 8) are C-occupied in 0 of 8, H3C15 in 1 of 8, and SPAG7 in 3 of 8 - C-occupied in every sample where it is N-occupied. The canonical KMT2A-fusion targets HOXA9 and MEIS1 are C-occupied in 0 of 8, establishing C-negativity as the expected signature of a fusion-bound locus.

    PMID 39472576DOI
    What this evidence cannot show

    These are counts out of 8 with no statistical test; SPAG7's 3-of-3 and IRX1's 0-of-3 are each based on three N-occupied samples, and the difference between them could arise by chance at that n. C-negativity at a locus could also reflect epitope accessibility rather than absence of wild-type MLL1, which is not excluded for any individual gene. HOXA9 and MEIS1 are used as positive controls for the fusion signature but were not pre-registered as such.

    Evidence that argues against this (2)
    • IRX1 is N-occupied in only 3 of 8 AFF1 samples, so '0 of 8 on the C arm' is being asked of a gene that the N arm itself mostly does not call. The comparison that matters is C-occupancy within the 3 samples where N fires, and there the denominator is three - which is also true of SPAG7, where it went the other way. Both calls rest on n = 3.
    • Removing SPAG7 and discounting H3C15 leaves a class defined by the three genes the authors' caller refutes plus IRX1. That is not a class surviving scrutiny, it is a single gene with a supporting cast that has been eliminated one member at a time across five cycles - and a result reached by serial elimination invites the question of whether the elimination criteria were applied evenhandedly to the survivor.
Connected to